Time-calibrated phylogeny of the woody Australian genus Hakea (Proteaceae) supports multiple origins of insect-pollination among bird-pollinated ancestors^†

Taxonomic sampling

We created a set of molecular data consisting of four regions from the DNA of the chloroplasts and other plastids (cpDNA) and three nuclear DNA (nDNA) regions from up to 75 species. Data for 70 of these species are presented here for the first time; data from the rest first were presented by 48 or 49, 50). The 75 species include 55 (of 149) species of Hakea, one species from each of the other 11 genera in tribe Embothrieae, one species from each of the other three tribes in subfamily Grevilleoideae, one species from each of the other four subfamilies in Proteaceae, and one species from each of the other two families of Proteales (Appendix 1; following 75). Our sampling of Hakea was taxonomically even, drawing 1–6 representatives from each of the 31 informal groups recognized by 5. 30 assigned species to bird- or insect-pollination classes on the basis of stigma–nectary distances (SND; >13 mm for the former, <13 mm for the latter). According to this diagnostic, we conclude that we included both insect- and bird-pollinated representatives from the Ulicina group of 5 but only the more common pollinator classes from three other groups with species that fall into both classes (Lorea, Varia, and Multilineata; one species in each of these falls into the alternate class and was not sampled).

DNA extraction, amplification, cloning, and sequencing

We extracted genomic DNA using the DNeasy Plant Mini Kit (Qiagen, Valencia, California, USA). The four cpDNA regions we sampled were the matK, atpB, and ndhF genes and the rpl16 intron; the three nDNA regions were waxy loci 1 and 2 and the PHYA gene. We amplified and sequenced all the DNA regions except the rpl16 intron, following the procedures used by 50. Following 50, we cloned all the nDNA regions but not the cpDNA regions. We sequenced the rpl16 intron following procedures described by 49.

Defining substitution characters

We aligned and edited the DNA sequences using Sequencher software (Sequencher version 4.7 sequence analysis software, Gene Codes Corporation, Ann Arbor, MI USA). We compared the plastid sequences to the complete plastid sequence of Nicotiana tabacum L. (GenBank accession number NC 001879) to determine the beginning and ending point of each cpDNA region. We compared the waxy sequences to the complete waxy sequence for Solanum tuberosum (GenBank accession number X58453) to determine the beginning and ending points of introns and exons and the numbering scheme for the exons. Alignment of all the coding regions was straightforward and did not imply any changes in reading frames, only a few insertions or deletions that were multiples of three nucleotides in length. We aligned the rpl16 and waxy intron sequences using the criterion of similarity, without regard to optimizing some description of the inferred phylogeny (66). We excluded from the data rpl16 and waxy intron sequences from outside of subtribe Hakeinae because we considered them to be too divergent across order Proteales to be confident of the homology of aligned positions. For modeling of nucleotide substitutions in the data, we recognized each of the plastid regions, the PHYA gene, the waxy locus 1 exons, the waxy locus 1 introns, the waxy locus 2 exons, and the waxy locus 2 introns as separate data partitions.

Coding morphological characters

We (R.M.B. and W.R.B.) modified the original morphological data of 59 characters for Hakea (6), adding taxa not previously included. For the present analysis, we simplified two transformation series from those data: one for leaf shape and division was simplified from four characters to two (characters 5 and 6, Table 1), and one for pollen presenter shape was simplified from two characters to one (character 29, Table 1). Analyses with and without use of these transformation series produced very similar results. Of the remaining 56 characters, 46 are parsimony informative and were used in our study (table 2).

Inferring phylogeny

We used the Akaike information criterion (Akaike, 1974) in MrModelTest 1.1b (available from J. A. A. Nylander, Uppsala University, Sweden, http://www.abc.se/∼nylander/) to select an adequately parameter-rich model of nucleotide substitution for each of the DNA regions. For the morphological data, we used the standard discrete model of MrBayes 3.1.2, which is based on the ideas of 39. These models then were used for their respective partitions in the Bayesian analyses in MrBayes (63). We unlinked the sampling of state frequencies, substitution rates, the gamma shape parameter, and the proportion of invariant sites for each DNA partition, and we linked branch lengths. We had MrBayes spawn two Markov Chain Monte Carlo (MCMC) runs when each partition was analyzed separately, and each of these MCMC runs ran four chains for 10 × 10⁶ generations. The program spawned three MCMC runs that each ran four chains for 20 × 10⁶ generations when the two partitions (the introns and the exons) of a waxy locus were analyzed together and four MCMC runs that each ran 12 chains for 40 × 10⁶ generations when more than two partitions were analyzed together. MrBayes sampled every 1000th generation in all analyses. All MrBayes analyses were run in parallel on the Florida State University High Performance Computing Cluster. These numbers of MCMC chains and generations resulted in potential scale-reduction factors (25) for all parameters within at least two units of 1 (i.e., <3; and typically within 0.1 of 1). We constructed a majority-rule consensus of the trees sampled after the burn-in in PAUP* v4.0b10 (69) to produce a tree where the posterior probabilities were ≥95% (or ≥50%, when morphological characters were analyzed alone) for internal nodes.

For comparison, we also ran a maximum-parsimony search for the morphological data in PAUP* (69). PAUP* performed a full heuristic search, tree-bisection-reconnection branch swapping, and 10 random-addition replicates. Because this resulted in more trees than could be stored in memory, we saved 100000 of the most parsimonious trees, created a strict consensus, and ran another search, using the same heuristic search settings, looking for equally parsimonious trees that were incongruent with this strict consensus. The morphological data also were bootstrapped in 1000 replicates by the same search strategy as for the maximum-parsimony search but with a maximum number of trees set at 10000.

Inferring ages

We used Thorne's Bayesian method (72; 71) to infer the age of divergence events using the six coding regions because the taxa outside subtribe Hakeinae were not represented in the intron partitions. Thorne's approach relaxes the assumption of a constant rate of nucleotide substitution over time by modeling rate change with temporal autocorrelation (26). We explored the robustness of our results to alternative prior distributions for the rate-change parameter ν (described by brownmean and brownsd) and the rate at the root node (described by rtrate and rtratesd), as described later. 65 demonstrated a strong correlation between the ages of divergence events inferred by Thorne's method and penalized likelihood (64) and the uncorrelated log-normal method implemented in BEAST (19), respectively, in a family level study of the Proteaceae.

We used 81 baseml program in PAML version 3.13 and Thorne's estbranches and multidivtime programs in the way described by 50. The tree topology inferred from the combined molecular and morphological data (the “total-evidence tree”; 36) was used for the final step with multidivtime. We used a mean of 1.15 (representing 1.15 × 10⁸ yr) for the prior distribution of time separating the most recent common ancestor (MRCA) of the ingroup (for this purpose, that of Platanaceae and Proteaceae) from the present (rttm). It is the mean of two age estimates for that node in a broad-scale dating study of the eudicots by 2. The standard deviation of this prior distribution (rttsd) was set to 0.15, a span that includes the earliest fossils that can be assigned to the stem of the Platanaceae, as described later. We used a mean of 0.0425 for the prior distribution of the rate of molecular evolution at the ingroup root node (rtrate). This value is one-half of the median distance between Plantanus L. and each taxon in the sister group (0.0978), as determined with the uncorrected-p parameter in PAUP*4.0b10 (69), divided by our prior estimate of time between the root of the ingroup and the present (1.15). The standard deviation of this prior distribution (rtratesd) was set to 0.0425 (equal to the mean; see 78, for a justification of making the standard deviation of this prior large relative to the mean). The “bigtime” parameter, which specifies what we consider to be a maximum age for the ingroup (in this case, a group composed of members of the Proteaceae and Platanaceae), was set at 1.3, equivalent to 130 Ma, older than the first appearance of the eudicots in the fossil record (45). The mean (brownmean) and standard deviation (brownsd) for the Brownian motion constant ν were both set to 1.0, as was the prior for the times of the interior nodes given the time of the root, according to recommendations made by Thorne in the multidivtime readme file. We left other parameters dealing with proposals in the MCMC at their default values (newk = 0.1, othk = 0.5, thek = 0.5).

Three constraints were used in the analyses. The maximum age for the MRCA of Proteaceae and Platanaceae was set at 125 Ma, the approximate time of appearance in the fossil record of the tricolpate pollen of the eudicots (45), of which Proteales is an early diverging branch. It has been used elsewhere as a calibration for the MRCAs of the eudicots (e.g., by 7), the Proteales and Sabiaceae (65), and the Proteaceae and Platanaceae (50). The minimum age for the MRCA of the Proteaceae and Platanaceae was set at 99.6 Ma, the upper boundary of the Albian (27), the earliest period from which fossil inflorescences have been recovered that can be assigned to Platanaceae (see, e.g., 10) on the basis of shared derived features (reviewed by 11; 2). It is slightly older than the occurrence of the oldest fossil that can be assigned to the Proteaceae from its position in parsimony analyses (65): Triorites africaensis (74; 16) from the Upper Cenomanian to Turonian (ca. 94 Ma) of Senegal and Gabon. The minimum age for the MRCA of Embothrium J.R.Forst. & G.Forst. and its sister, Telopea R.Br., was set to 35.4 Ma on the basis of the occurrence of Granodiporites nebulosus at that time (41; 65). Granodiporites nebulosus was resolved as sister to Embothrium by 65. 4 and 65 both used this fossil calibration in their studies of the family. 4 used a second calibration within tribe Embothrieae (the fossil pollen species Propylipollis ambiguus), but we do not use it here because 65 demonstrated that it cannot be assigned such precision with confidence. 65 also demonstrated that Hakeidites martinii, a pollen fossil previously assigned to Hakea by 35, cannot be assigned this precision with confidence.

We explored the sensitivity of the inference of ages to assumptions made in the dating analysis by systematically varying the combinations of the rtrate values of 0.00425, 0.0425 (the original value), and 0.425 and the brownmean values of 0.1, 1 (the original value), and 10; the standard deviation of each prior was set equal to the mean. A similar set of sensitivity tests was performed by 78, 7, and 50. We also determined the effect of removing the minimum age constraint for the MRCA of Embothrium and Telopea in one analysis and, in another, the minimum and maximum age constraints for the MRCA of Proteaceae and Platanaceae. In a final test of sensitivity, we increased the bigtime parameter to 2.6 and 13, in turn, while using the original values for rtrate and brownmean.

Coding pollen vectors

30 examined the distribution of SNDs from 25 species of Hakea for which floral visitors have been observed or postulated. They found all known or putative insect-pollinated species to have SND <13 mm and all known or putative bird-pollinated species to have SND >13 mm. They presented evidence of why SND is predictive by noting that the shortest bill length of the principal bird pollinators in Australia (the honeyeaters, family Meliphagidae) is 12 mm (58) and that insect pollinators in Western Australia rarely exceed a body length of 15 mm. We took SND from 30 for species included in that study; approximated SND for most of the remaining species from straightened pistil lengths given by 5, taking into account whether the pistil is curved; and measured SND from herbarium specimens when a species’ pistil was slightly longer than 13 mm and known to be curved. Although both the bill length of the honeyeaters and body length of insect pollinators might have differed considerably in the past from today's size ranges, 30 demonstration of a significant evolutionary correlation between inflorescence color (red is often, though not exclusively, associated with bird pollination) and SND size class in Hakea suggests that SND is a good predictor of pollinator class throughout the history of the genus.

Inferring ancestral states

We used two methods to determine ancestral states on the phylogeny. The Ancestral State Reconstruction Packages for Mesquite (43) in Mesquite (44) inferred character state changes from 22 parsimony on the chronogram inferred on the basis of our preferred set of prior assumptions. In the other method, MultiState in BayesTraits (57) calculated the maximum likelihood of the data when the MRCA of Hakea in the same chronogram was held as one state, then the other, using the “fossil” command. We used a log-likelihood difference of at least 2 as our threshold of significance, which is a rule of thumb used in the same way elsewhere (e.g., 56), following 20. We used a two-parameter model in which the evolutionary rates between the two states could be unequal because the log-likelihood of the data was significantly better (P = 0.0144) when we used that model (–28.526399) compared with the one-parameter model of equal rates (–31.519877), as determined using a likelihood ratio test with 1 degree of freedom. We excluded the outgroup taxa in these analyses to avoid introducing biases in estimates of evolutionary rate that might occur when a single taxon is used to represent a much larger group (as discussed, e.g., by 52). BayesTraits cannot calculate likelihood values for trees with polytomies, so we had Mesquite randomly resolve the two polytomies in the chromatogram and give the new branches a length of 0.00001 (the minimum length accepted by BayesTraits; the equivalent of 1000 yr).

Tests of evolutionary correlation, ordered evolution, and trait contingency

Because our results suggested that the morphology-based phylogeny of Hakea by 6 should be rerooted, we took the additional step to reanalyze the data on pollinators and floral cyanide from 30 following the same approach (from 56) and protocol of analysis that they did using the BayesDiscrete part of BayesTraits (57), with the original rooting and the new rooting. Specifically, we used the data and alternative rootings to determine whether evidence exists for a correlation between SND and floral cyanide concentrations, for the contingency of the evolution of bird pollination on high floral cyanide levels, and for greater rates from insect pollination + low cyanide concentrations to insect pollination + high cyanide concentrations than from the former to bird pollination + low cyanide concentration. Because their Table S1 did not include H. clavata Labill. or H. megalosperma Meisn., we excluded these from our analyses. 30 considered mean values for their floral cyanide assay ≤5 Feigl-Anger units (FA) to represent low cyanide concentrations and mean FA values ≥7.5 to represent high concentrations. The FA values for three species, H. erecta Lamont (mean = 5.5), H. recurva Meisn. (5.4), and H. smilacifolia Meisn. (6.8), fall between these thresholds. For our analyses, we considered the first two to have low floral cyanide concentrations and the last to have high concentrations. We performed the analyses with branch lengths equal (all set to 1), as done by 30. For the analysis, we used the topology shown in Figure 3 of 30. 30 referred to analyses performed on 10 trees that represented the most divergent possible resolutions of polytomies in the 6 topology. We did not attempt to recreate the other nine trees as well because our interest was not in exhaustively reanalyzing 30 results but rather in whether a rerooting of 6 topology made a difference.

Defining substitution characters

The sampled DNA regions are described in Table 3. We generated up to 9322 aligned positions for 75 taxa—ca. 700 kilobases of data in total. Each of the 75 species is represented in each set of coding DNA (which excludes the rpl16 and waxy introns) with seven exceptions (Appendix 1). Approximately 75% of the sites in the waxy intron alignments could be unambiguously aligned across subtribe Hakeinae; the remaining intron sites were not used in our study. Alignments of the genes matK, atpB, and ndhF represent the complete sequences for those genes without bordering regions. The alignment of the rpl16 intron does not include 36 nucleotides from the 5′ end and seven nucleotides from the 3′ end of the intron. Alignment of waxy exons 7–10 includes only 24 nucleotides on the 3′ end of exon 7 and only 171 nucleotides on the 5′ end of exon 10; exons 8 and 9 are complete. The DNA alignments are available as Appendix S1 (see Supplemental Data with the online version of this article).

Coding morphological characters

Table 1 shows the morphological characters sampled, and table 2 shows the character-state data for each taxon. Including the three additional species of Grevillea used by 6; G. anethifolia R.Br., G. glauca Banks & Sol. ex Knight, and G. hakeoides Meisn.) in the morphological analysis did little to change the results, and because they were not sampled in the DNA data, they are not included in the analyses presented here.

Inferring phylogeny

Table 3 shows the models chosen as adequately parameter-rich for each set of data. The trees inferred from the individual cpDNA regions showed no topological incongruities (Appendix S2). Fig. 1 presents the trees for the four cpDNA regions combined and the individual nDNA regions; Fig. 2 presents the results of the parsimony and Bayesian analyses of the morphological data. Few incongruities exist between each of these topologies and the total-evidence topology (Fig. 3): the combined cpDNA topology is incongruous in four places; the waxy locus 1 topology in four; the waxy locus 2 topology in six; the PHYA topology in eight; the parsimony result from the morphological data in one; and the Bayesian result from the morphological data in 10.

The great number of incongruities with the PHYA topology arises because of an unexpected clade composed of members from both clades formed at the basal split of the total-evidence tree (Clades A and B; Fig. 3). The other two clades that share a polytomy with this clade in the PHYA topology are otherwise identical to Clades A and B, and they are labeled as such using quotation marks. Three processes are generally recognized as leading to incongruity among gene (or “linkage partition,” such as is the case for cpDNA) trees (53; 17; 42; 68): (i) lateral gene transfer, including hybridization and introgression; (ii) gene duplication (which can lead to paralogous comparisons); and (iii) deep coalescence (i.e., lineage sorting). We were unable to determine the cause of this incongruity: hybridization is unknown among extant populations of the six species in the anomalous clade; we did not observe two bands in our gel electrophoreses of PHYA for any species of Hakea; we did not find a second locus in the 10 colonies that we sequenced from the cloning of each of the six species involved; and the branch lengths between H. lorea and H. multilineata are not particularly long, as might be expected if they represent a period of time that includes the basal divergence in Hakea. Because the relationships inferred in the combined DNA and total-evidence trees did not depend on inclusion of this anomalous clade (but branch lengths did), we excluded the PHYA data for these six species in the combined DNA and total-evidence analyses (following 38).

The basal split in the morphology-based results is between H. lorea and the remaining sampled species. Although this rooting is congruent with the results of 6, the other early branching events are not. Few branches in the morphology-based results were well supported though (e.g., there was just one branch with bootstrap support >70% and two with posterior probabilities >90%), and all but one of the 10 incongruities between the Bayesian result and the total-evidence phylogeny had <90% posterior probabilities in the former. 6 did not determine support values for branches, so we cannot compare the results of the two studies in this regard.

The total-evidence topology (Fig. 3) is identical to that found when the seven DNA regions were analyzed in combination without the morphological data. Of the 19 informal groups of 5 from which two or more species were sampled in our study, 12 are resolved as nonmonophyletic and seven as monophyletic in the total-evidence tree (Fig. 3). Of the five informal groups of 5 used by 30 to identify closely related species for comparison (Hanley et al. table 2; their species group “Sulcata” is not separate from the Ulicina group in 5, and so is not counted here), four of them are not monophyletic in our total-evidence tree. Five of the six multigroup clades named by 6 are not monophyletic on the phylogeny from our analysis of molecular and morphological data combined, but the largest of these, the Multilineata clade, is monophyletic and makes up most of Clade B (Fig. 3).

Inferring ages

We inferred the age of the crown group of Hakea to be 9.6 Ma (= mean; 95% credible interval = 6.4–14.0) using our preferred set of parameter values and all three age constraints. The mean inferred age varied from 8.1 Ma to 17.4 Ma as we varied values of rtrate, rtratesd, brownmean, and brownsd. The highest age estimates were inferred when the prior of the Brownian motion constant ν was set to 0.1; this is the most “clock-like” value assumed for that variable, an assumption that is perhaps inappropriate given the distribution of branch lengths (see phylogram, Fig. 4c). The mean of all other age estimates for the crown group was <10 Ma. For example, removing the minimum age constraint (35.4 Ma) for the MRCA of Embothrium and Telopea resulted in a mean inferred age of 7.7 Ma (4.8–11.3) for the crown group of Hakea, and removing the minimum and maximum age constraint for the MRCA of Proteaceae and Platanaceae resulted in a mean inferred age of 9.6 Ma (6.3–13.8). Increasing the bigtime parameter did not substantially change the ages inferred by the method. The second replicate for each of these analyses resulted in mean and boundaries of the 95% credible interval values within 0.3 Ma of those of the first replicate. Using our preferred set of parameter values and all three age constraints, we inferred the age of the stem group of Hakea to be 12.8 Ma (8.8–18.0).

Coding pollen vectors

Of 19 species included in both our study and that of 30, nine have SND <13 mm and 10 SND >13 mm. Of the 36 species for which we needed to estimate SND from pistil lengths from 5, 27 had pistil length <13 mm and nine had pistil length >13 mm. Of those with pistil length >13 mm, two included 13 mm in the bottoms of their ranges: H. pandanicarpa R.Br. (13–16 mm) and H. grammatophylla F.Muell. (13–16 mm). The flowers of H. pandanicarpa have a strongly incurved pistil. We measured a mean SND of 11.9 mm (range 9–11.5 mm) for H. pandanicarpa from a single flower from each of five specimens at the National Herbarium of New South Wales (NSW). We consider this result and the strong, sweet scent of its flowers (82) to be evidence that it is insect pollinated. The flowers of H. grammatophylla have a bent, but not strongly incurved, pistil. We measured a mean SND of 14.9 mm (range 13–17) for H. grammatophylla from a single flower from each of five specimens at NSW. We consider this and its deep pink flowers in large, exposed inflorescences to be evidence that it is bird pollinated. These codings of pollinator for the species are consistent with those proposed by 6 in cases of overlap with that study, with the exception that 6 considered the Corymbosa group (represented by H. victoria and H. corymbosa here) to be mammal, rather than bird, pollinated (contra 30).

Inferring ancestral states

Parsimony inference of ancestral SNDs resolve the MRCA of Hakea to have been bird pollinated (to have had SND >13 mm) with either nine shifts to insect pollination (SND <13 mm) or seven shifts to insect pollination and two back to bird pollination (Fig. 4b). BayesTraits calculated the rate of change from bird to insect pollination is 10.752508, and the rate of change from insect to bird pollination is 0. The log-likelihood of the data is significantly different when the MRCA of Hakea is held as bird pollinated (–28.526399) as opposed to insect pollinated (–31.447590).

Tests of evolutionary correlation, ordered evolution, and trait contingency

Using 30 original data and tree (the topology of their Fig. 3), we found both the correlation between SND and floral cyanide concentrations (χ² = 12.92, df = 4, P = 0.012) and evidence that the evolution of bird pollination (SND >13 mm) is contingent on high cyanide levels (contingent evolution; χ² = 5.04, df = 1, P = 0.025) to be significant, but the evidence for ordered evolution (a greater evolutionary rate from insect pollination + low cyanide concentrations to insect pollination + high cyanide concentrations than to bird pollination + low cyanide concentrations; χ² = 0.000002, df = 1, P = 0.999) not to be. Oddly, the P value we found for the test of ordered evolution (P = 0.999) was not within the range that 30 report for the test with each of the 10 tree topologies (P = 0.03–0.20), but our results for the correlated evolution and contingent evolution tests are similar to theirs. After rerooting their topology to make the ingroup rooting consistent with what we found, evidence for correlated evolution between the two characters remained significant (χ² = 12.54, df = 4, P = 0.014), and evidence of ordered evolution remained nonsignificant (χ² = 0.000002, df = 1, P = 0.999), but the evidence for contingent evolution became less significant (χ² = 2.31, df = 1, P = 0.129).

Phylogenetic congruence with previous taxon circumscriptions

We inferred that the basal split in Hakea produced two clades (Fig. 3), each with moderate morphological cohesiveness. We estimate that the larger clade (Clade A) contains 96 species, on the basis of memberships in the informal groups of 5. All sampled members of each informal group, whether it is monophyletic or not, appear in either Clade A or Clade B but not both. Clade A is characterized by obscure leaf venation and discoid pollen presenters, with the exception of the Lissocarpha and Varia groups, which have conical pollen presenters. We estimate that Clade B contains about half as many species (53) as Clade A. It is composed primarily of the Multilineata clade recognized by 6; Fig. 3) and is characterized by prominent leaf venation and conical pollen presenters. Determination of the derived state in each of these two characters will have to await clarification regarding the sister group to Hakea, since it might not be the whole of Grevillea and Finschia. The two clades formed at the basal split have not been recognized previously.

Other putative multigroup clades of 6 and 12 of the 19 informal groups of 5 sampled multiple times in our study are nonmonophyletic in the total-evidence phylogeny (Fig. 3) and thus should not be considered monophyletic in future evolutionary studies. Although this result at first could be viewed as a striking incongruity with 5 study, we note that five of the 12 were not resolved as monophyletic in the cladogram of 6 either and that four other nonmonophyletic informal groups were represented by only one species in the 6 study, so their monophyly was not previously tested. Although branch-support values were not reported by 6, those we present (Fig. 2) suggest that the morphological data alone do not produce strong corroboration for the monophyly of the three remaining informal groups (relevant bootstrap values all <50%), so the incongruence is not worrisome.

The age of Hakea

The recent origin of the crown group of Hakea (ca. 10 Ma) is unexpected on the basis of both 30 expectations and comparisons with the ages of ecologically similar lineages. For example, in a study of diversification timing in four sclerophyll lineages in Australia, 12 found that much of the extant diversity of those groups arose from rapid diversification throughout an earlier period that corresponds mostly to the early and mid Miocene (25–10 Ma ago, according to 12). At that time, Australia's climate is thought to have become drier and more seasonal, as marked by a decrease, followed by disappearance, of Nothofagus pollen and a marked increase in pollen of Myrtaceae and other sclerophyll taxa in the fossil record (47). Although good phylogenetic evidence exists for the evolution of adaptations to fire as early as the Cretaceous–Palaeogene boundary (14), fire is thought to have become more regular in the late Miocene (47), a period between 11.6 and 5.3 Ma ago (27) that provides a reasonable context for the increase in follicle-wall thickness that is a synapomorphy for the genus Hakea. This conclusion is consistent with 6 view that the Miocene would have been a time conducive to the radiation of Hakea, had this lineage originated by then.

Our Bayesian age estimates were robust against large variations in our prior assumptions, and our age estimate for the stem group of Hakea (mean = 12.8 Ma, 95% credible interval = 8.8–18.0 Ma; Fig. 4a) is similar to that found in the family-wide study by 65; mean = 14.1 Ma, 95% credible interval = 8.3–21.52 Ma) when they also used Thorne's Bayesian method (72; 71). Using a method in which autocorrelation in rates in adjacent branches is not assumed (19), they estimated the stem age to be only slightly older (mean = 16.2 Ma; 95% credible interval = 9.2–23.3 Ma). 65 use of a single OTU for Hakea means that they did not infer the age of its crown group. Another recent, family-wide dating study (4) did not include a representative of Hakea. If Grevillea proves to be paraphyletic and we failed to sample the sister lineage of Hakea in the current study, then our age estimate for the stem group of Hakea will prove to be higher, not lower, than justified.

The evolution of pollination

On our phylogeny of Hakea, parsimony and likelihood approaches reconstructed the MRCA of the genus to have had an SND that occurs in extant species that use birds as pollinators (>13 mm), rather than one that corresponds to insect pollination (<13 mm) as found in 30 more restricted geographic and taxonomic sampling. The principal bird pollinators of Hakea, the honeyeaters (6; 30), appear to have been present in Australia from at least the Miocene (8), making reasonable the inference of bird pollination in a 10-Ma-old MRCA of Hakea.

Our parsimony inference of multiple shifts from bird to insect pollination, rather than the reverse, and our maximum likelihood inference of a rate of 0 from insect to bird pollination (as opposed to 10.8 for the reverse) make Hakea only the third documented example of this kind of transition that we can find in the literature on pollinator shifts. However, that literature is heavily biased toward New World examples, which have focused on taxa in which the only bird-pollinated plants are those associated with hummingbirds. Ours is only the third phylogenetic study that has examined transitions between insect and honeyeater pollinators, all of which have raised the possibility of shifts from bird to insect pollination. More phylogenetic results like ours are required before such a pattern can reasonably be said to demand explanation, but striking differences in the foraging behavior of these two bird clades suggest that such a pattern might be expected. Hummingbirds, which hover when feeding, can be expected to show a finer degree of precision in their interactions with their food plants than honeyeaters, which perch to feed. Moreover, honeyeater-pollinated plants tend to be visited by a wide range of bird species of dramatically variable size, which themselves feed from a wide range of plant species (23). Even species of Hakea from Australia's arid center are exposed to a fauna of 11 meliphagid species, ranging in length from 10 to 28 cm (67). This lack of specialization compared with hummingbird-pollinated plants implies the possibility of greater flexibility in shifting to generalist pollination and thence to more specialized insect pollination than for hummingbird-pollinated taxa.

Our inability to replicate 30 “ordered evolution” and, with a rerooting of the ingroup, “contingent evolution” results suggest that 30 conclusion that bird pollination only arose where insect-pollinated ancestors were already producing high concentrations of floral cyanide is incorrect. That said, new attention might be productively shifted to groups diverging earlier in the phylogeny than the crown group of Hakea because the bird pollination primitive to the crown group can be assumed to have arisen at some time leading up to it, and neither our study nor that of 30 examined that pollination shift. 23 estimated that 150 of the 357 species of Grevillea are bird pollinated.

The divergence events leading up to the MRCA of Hakea and the species of Grevillea sampled in our study might reasonably be expected to have produced other lineages of Grevillea, given that our results and the monophyly of Grevillea would otherwise imply that the diversification that produced the third largest genus in the Australian flora (357 species) occurred over a period of just 13 Ma. A young age for Grevillea would be at odds with the results of 40, who, in a review of published studies of clade ages, concluded that the Australian flora appears to consist largely of slowly diversifying, ancient lineages (>15 Ma old). On the other hand, it would support the hypothesis of 32, 33) that a period of slower speciation was followed, in the late Tertiary and Quaternary, by an acceleration of speciation in southwestern Australia (where large fractions of Hakea and Grevillea are endemic). Clearly, further sampling of Grevillea (Olde and Marriott 1995; Makinson 2000) is needed before hypotheses regarding the evolutionary interplay between pollination and florivory in subtribe Hakeinae (a now justifiably broader focus than just Hakea) and diversification rates in the subtribe can be adequately addressed.